{
  "id": 497103,
  "title": "Rookie Needs Help",
  "url": "/competitions/birdclef-2024/discussion/497103",
  "author_name": "",
  "post_date": "2024-04-23T16:21:28.983196600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hi all, I created a simple neural network for this competition, but when I use the model to do prediction, it produces the same predictions regardless of input.</p>\n<p>So far I've confirmed that the training images are normalized to have values between 0 and 1 and the training labels are properly made into categoricals.  The model itself is simple, maybe too simple?</p>\n<p>I'm assuming the problem is somewhere in the model itself, so here's a link to the notebook:<br>\n<a href=\"https://www.kaggle.com/osprey2k/birdclef-2024-model1\" target=\"_blank\">https://www.kaggle.com/osprey2k/birdclef-2024-model1</a></p>\n<p>Any and all help is appreciated!</p>",
  "messages": [
    {
      "id": "2770000",
      "postDate": "04/23/2024 16:21:28",
      "content": "<p>Hi all, I created a simple neural network for this competition, but when I use the model to do prediction, it produces the same predictions regardless of input.</p>\n<p>So far I've confirmed that the training images are normalized to have values between 0 and 1 and the training labels are properly made into categoricals.  The model itself is simple, maybe too simple?</p>\n<p>I'm assuming the problem is somewhere in the model itself, so here's a link to the notebook:<br>\n<a href=\"https://www.kaggle.com/osprey2k/birdclef-2024-model1\" target=\"_blank\">https://www.kaggle.com/osprey2k/birdclef-2024-model1</a></p>\n<p>Any and all help is appreciated!</p>",
      "rawMarkdown": "Hi all, I created a simple neural network for this competition, but when I use the model to do prediction, it produces the same predictions regardless of input.\n\nSo far I've confirmed that the training images are normalized to have values between 0 and 1 and the training labels are properly made into categoricals.  The model itself is simple, maybe too simple?\n\nI'm assuming the problem is somewhere in the model itself, so here's a link to the notebook:\n[https://www.kaggle.com/osprey2k/birdclef-2024-model1](https://www.kaggle.com/osprey2k/birdclef-2024-model1)\n\nAny and all help is appreciated!",
      "votes": null
    },
    {
      "id": "2770096",
      "postDate": "04/23/2024 17:20:27",
      "content": "<p>As a fellow rookie in audio, I'll try my best here. I'm assuming (maybe incorrectly) that the model might be too simple and it's deciding to just settle in a local minima across all bird species. Since there are 182 labels, I'm wondering if that same output you're seeing is like a uniform distribution on all classes. Model might be like \"if I guess the same Prob for all birds, I can feel same in my loss function blanket\" and such. </p>\n<p>Maybe make the model more complex or try combining audio files. I've been playing around with mixing 3-6 files together so it can learn more \"realistically\" for the test time data.</p>\n<p>Curious to see more seasoned people answer here, but that's my educated/uneducated guess.</p>",
      "rawMarkdown": "As a fellow rookie in audio, I'll try my best here. I'm assuming (maybe incorrectly) that the model might be too simple and it's deciding to just settle in a local minima across all bird species. Since there are 182 labels, I'm wondering if that same output you're seeing is like a uniform distribution on all classes. Model might be like \"if I guess the same Prob for all birds, I can feel same in my loss function blanket\" and such. \n\nMaybe make the model more complex or try combining audio files. I've been playing around with mixing 3-6 files together so it can learn more \"realistically\" for the test time data.\n\nCurious to see more seasoned people answer here, but that's my educated/uneducated guess.",
      "votes": null
    },
    {
      "id": "2770355",
      "postDate": "04/23/2024 19:48:18",
      "content": "<p>Thanks Matt!  I dug a bit deeper at your suggestion and found that batch normalization seems to help with this problem.  I implemented some batch normalization layers and found that now I am getting different predictions using my model.</p>",
      "rawMarkdown": "Thanks Matt!  I dug a bit deeper at your suggestion and found that batch normalization seems to help with this problem.  I implemented some batch normalization layers and found that now I am getting different predictions using my model.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2770096,
      "author_name": "msthil",
      "author_url": "",
      "post_date": "04/23/2024 17:20:27",
      "content": "<p>As a fellow rookie in audio, I'll try my best here. I'm assuming (maybe incorrectly) that the model might be too simple and it's deciding to just settle in a local minima across all bird species. Since there are 182 labels, I'm wondering if that same output you're seeing is like a uniform distribution on all classes. Model might be like \"if I guess the same Prob for all birds, I can feel same in my loss function blanket\" and such. </p>\n<p>Maybe make the model more complex or try combining audio files. I've been playing around with mixing 3-6 files together so it can learn more \"realistically\" for the test time data.</p>\n<p>Curious to see more seasoned people answer here, but that's my educated/uneducated guess.</p>",
      "votes": null,
      "replies": [
        {
          "id": 2770355,
          "author_name": "osprey2k",
          "author_url": "",
          "post_date": "04/23/2024 19:48:18",
          "content": "<p>Thanks Matt!  I dug a bit deeper at your suggestion and found that batch normalization seems to help with this problem.  I implemented some batch normalization layers and found that now I am getting different predictions using my model.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2770000": "Hi all, I created a simple neural network for this competition, but when I use the model to do prediction, it produces the same predictions regardless of input.\n\nSo far I've confirmed that the training images are normalized to have values between 0 and 1 and the training labels are properly made into categoricals.  The model itself is simple, maybe too simple?\n\nI'm assuming the problem is somewhere in the model itself, so here's a link to the notebook:\n[https://www.kaggle.com/osprey2k/birdclef-2024-model1](https://www.kaggle.com/osprey2k/birdclef-2024-model1)\n\nAny and all help is appreciated!",
    "2770096": "As a fellow rookie in audio, I'll try my best here. I'm assuming (maybe incorrectly) that the model might be too simple and it's deciding to just settle in a local minima across all bird species. Since there are 182 labels, I'm wondering if that same output you're seeing is like a uniform distribution on all classes. Model might be like \"if I guess the same Prob for all birds, I can feel same in my loss function blanket\" and such. \n\nMaybe make the model more complex or try combining audio files. I've been playing around with mixing 3-6 files together so it can learn more \"realistically\" for the test time data.\n\nCurious to see more seasoned people answer here, but that's my educated/uneducated guess.",
    "2770355": "Thanks Matt!  I dug a bit deeper at your suggestion and found that batch normalization seems to help with this problem.  I implemented some batch normalization layers and found that now I am getting different predictions using my model."
  },
  "source": "meta"
}